Computing 2D projections (or images) from a 3D dataset has found increased use in many applications, such as image guidance and image rendering. However, conventional methods usually bear a high computational complexity. New and improved methods for quickly computing 2D images from a 3D dataset are needed.
For example, image guidance is a daily practice in clinical institutions around the world. It provides unprecedented positioning accuracy and improved quality of treatment, vital to many clinical procedures such as biopsy, surgery, interventional procedures, and radiation treatments. For example, in image-guided radiotherapy (IGRT), a pre-operative image is usually a 3D computed tomography (CT) image used for treatment planning At treatment time, intra-operative images are acquired and compared with the pre-operative image to determine alignment errors. Compared with 3D modalities such as Cone Beam CT, 2D X-ray imaging is often chosen for intra-operative imaging due to its fast acquisition and low imaging dose, and 3D-2D image registration is usually employed to register the 3D pre-operative CT image with 2D X-ray intra-operative images acquired with patient on the treatment couch.
Since 3D and 2D images cannot be directly registered, the 2D X-ray intra-operative images are usually registered with synthetic images called DRRs. DRRs are simulated X-ray-like images created from a 3D CT image by placing the 3D volume (represented by that CT image) in specific poses within a virtual imaging system that simulates the actual X-ray imaging geometry.
Conventional registration methods usually operate iteratively: start with a hypothetical pose, create DRRs for that pose, evaluate similarity with intra-operative images, and search for the next potential pose; then generate DRRs for this new pose, and the iteration goes on. During the search, DRRs may be generated for multiple poses close to the current pose for numerical estimation of derivatives. For example, if the method employs a six-dimensional (6D) search (i.e., 3D translation and 3D rotation), additional DRRs will be generated for six close poses just for a gradient estimation at one pose. Typically DRRs will be generated for hundreds of times during registration.
DRR generation is computationally intensive in nature, and conventional methods usually bear a high complexity level. Given its complexity and frequency of execution during registration, faster DRR generation methods are highly desired.